Long time propagation and control on scarring for perturbed - - PDF document

Long time propagation and control on scarring for perturbed quantized hyperbolic toral automorphisms

Jean-Marc Bouclet∗ Stephan De Bi` evre† Universit´ e de Lille 1 UMR CNRS 8524, 59655 Villeneuve d’Ascq

Abstract We show that on a suitable time scale, logarithmic in , the coherent states on the two- torus, evolved under a quantized perturbed hyperbolic toral automorphism, equidistribute

• n the torus. We then use this result to obtain control on the possible strong scarring of

eigenstates of the perturbed automorphisms by periodic orbits. Our main tool is an adapted Egorov theorem, valid for logarithmically long times.

1 Introduction

One of the main results in quantum chaos is the Schnirelman theorem. It states that, if a quantum system has an ergodic classical limit, then almost all sequences of its eigenfunctions converge, in the classical limit, to the Liouville measure on the relevant energy surface [7, 15, 20, 24]. It is natural to wonder if the result holds for all sequences (a statement commonly referred to as “unique quantum ergodicity”). This has been proven to be true for the (Hecke) eigenfunctions of the Laplace-Beltrami operator of a certain class of constant negative curvature surfaces [17] and has been conjectured to be true for all such surfaces [19]. It also has been proven to be wrong for quantized toral automorphisms in [11]. In that case, sequences of eigenfunctions exist with a semiclassical limit having up to half of its weight supported on a periodic orbit of the dynamics. This phenomenon is referred to as (strong) scarring. In [5, 12], it is shown that this last result is

• ptimal: if a measure is obtained as the limit of eigenfunctions then its pure point component can

carry at most half of its total weight. Except for the Schnirelman theorem, which holds in very great generality, all cited results are proven by exploiting to various degrees special algebraic or number theoretic properties of the systems studied. It is one of the major challenges in the field to device proofs and obtain results that use only assumptions on the dynamical properties of the underlying classical Hamiltonian system, such as ergodicity, mixing or exponential mixing, the Anosov property, etc. without relying on special algebraic properties. It is argued in [4, 5, 12] for example, that this will require a good control on the quantum dynamics for times that go to infinity (at least) logarithmically as the semiclassical parameter

∗Jean-Marc.Bouclet@math.univ-lille1.fr †Stephan.De-Bievre@math.univ-lille1.fr

1

goes to zero: t ≥ k− ln for some constant k− > 0. It is well known that such control is in general hard to obtain especially since a good lower bound on k− is needed. In this paper, we concentrate

• n the quantized perturbed hyperbolic automorphisms of the 2d-torus, which are known to be

Anosov systems classically. For those systems, we first prove an Egorov theorem valid for times proportional to ln , with an explicit control on the proportionality constant k− (Theorem 3.1). This result is obtained by adapting the techniques of [8]. We then combine this result with recent sharp estimates on the exponential mixing of the classical dynamics [3] to study the long time evolution of evolved coherent states (Theorem 4.7), showing that on a sufficiently long logarithmic time scale, those evolved coherent states equidistribute on the torus. Roughly, the result is that for all f ∈ C∞(T2), Q(f, t, ) ≡

• U t

ǫϕa ,κ, OpW (f)U t ǫϕa ,κ

• H(κ) −
• T2 f(x) dx → 0,

→ 0, (1.1) for times k− ln ≤ t ≤ k+ ln , 0 ≤ k− ≤ k+. Here Uǫ is the unitary quantum dynamical evolution operator, OpW (f) is the Weyl quantization of f, and ϕa

,κ is a coherent state at the point a of the two-torus T2. For detailed definitions, we refer

to the following sections. This result generalizes results obtained in [4] for unperturbed hyperbolic

• automorphisms. To prove it, we prove an estimate of the type

Q(f, t, ) ≤

• ϕa

,κ, (U −t ǫ OpW (f)U t ǫ − OpW (f ◦ Φt ǫ))ϕa ,κ

• H(κ)
• +
• ϕa

,κ, OpW (f ◦ Φt ǫ))ϕa ,κ

• H(κ) −
• T2 f(x) dx
• ≤

ǫ1(eγqt) + ǫ2(−1e−γct). Here ǫ1 and ǫ2 are functions tending to zero when their argument does. The first term comes from the error term in the Egorov theorem, whereas the second one involves a classical mixing rate γc. It is obvious that this estimate leads to the result only if γq ≤ γc. One therefore needs γc to be large (fast mixing) and γq to be small. Sharp results on the classical mixing rates of Anosov systems are hard to come by, but for some Anosov maps, among which the perturbed toral automorphisms that are the subject of this paper, such results have become available recently [3]. The remaining difficulty resides therefore in controlling the exponent in the error in the Egorov theorem. This is dealt with in the next section. We note that, although we prove the Egorov theorem for systems on the 2d-torus, the result above is only proven for d = 1. Indeed, noting Γmin and Γmax for the smallest and largest Lyapounov exponents of the system, we prove in Section 2 that, essentially, γq = 3

2Γmax. On the

• ther hand, the available estimates on the classical mixing rate [3] yield essentially γc = 2Γmin. Of

course, when d = 1, Γmax = Γmin and we have γq ≤ γc as needed. For d > 1, on the other hand,

• ur proof of (1.1) still goes through, but only under an artificial “pinching” condition of the type

3Γmax < 4Γmin. As an application of the above result, we finally show how to use the information obtained on the evolved coherent states in combination with the basic strategy of [5, 6] to gain some control

• n the scarring of eigenfunctions (Theorem 4.9, Corollary 4.10). Roughly speaking, we show that

if a sequence of eigenfunctions of a quantized perturbed hyperbolic toral automorphism converges to a delta measure on a finite union of periodic orbits, then it must do so slowly. An improvement

• n this result (basically, on how slowly) has been announced recently in [13]. We don’t expect

this result to be optimal: indeed, it is expected, as in the case of unperturbed automorphisms, 2

that sequences of eigenfunctions can not concentrate completely on periodic orbits, no matter how

• slowly. Proving this would involve controlling the quantum dynamics for longer times than we are

currently able to do. A result somewhat analogous to our result on the evolution of coherent states was recently

• btained for the long time evolution of Lagrangian states on compact Riemannian manifolds of

negative curvature [21]. It should however be noted that such a result does not require any control

• n the proportionality constant preceding ln so that no precise control on either the mixing rate
• r the exponent in the error term of the Egorov theorem are needed in that case. We suspect that

in situations were such control can be obtained, our present strategy will allow to control both coherent state evolution and strong scarring. A related result for the eigenfunctions of Laplace-Beltrami operators on compact, negatively curved Riemannian manifolds is proven using a different strategy in [1]: it is shown there that (under a suitable technical condition that may or may not hold) such eigenfunctions can not concentrate on sets of small topological entropy (and therefore on periodic orbits).

2 Weyl quantization and Egorov Theorem

The purpose of this section is to recall (as compactly as possible) some properties of the Weyl quantization on T2d := (R/Z)2d as well as on R2d, for d ≥ 1. More specifically, we want to state a semi-classical version of the Egorov Theorem in the case of T2d. The latter is of course well known for R2d but it requires a proof for T2d all the more so as we need a rather explicit version of this theorem for the applications we have in mind in this paper. The Weyl quantization on R2d can be defined as the linear map f ∈ B(R2d) → OpW (f) ∈ L(L2(Rd)) where OpW (f) is the operator (belonging a priori to L(S(Rd), S′(Rd))) with Schwartz kernel Kf(q1, q2) = (2π)−d

• Rd exp(i(q1 − q2) · p) f

q1 + q2 2 , p

• dp.

Here B(R2d) is the set of smooth functions f on R2d such that ∂γf is bounded for all γ ∈ N2d, thus the above integral has to be understood in the sense of oscillatory integrals [16, 22, 18, 14] but it is

• f course a usual Lebesgue integral if f decays fast enough at infinity. The fact that OpW (f) can

be considered as a bounded operator on L2(Rd) follows from the Calder`

• n-Vaillancourt theorem

[16, 22, 18, 14] which states the existence of C > 0 and ¯ d > 0 such that

• OpW (f)ψ
• L2(Rd) ≤ C sup

|γ|≤ ¯ d

||∂γf||L∞(R2d)||ψ||L2(Rd), ∀ f ∈ B(R2d), ψ ∈ S(Rd). (2.2) It is moreover well known that OpW (f) maps the Schwartz space S(Rd) continuously into itself and that OpW (f)∗ = OpW (f), thus OpW (f) can be considered as a continuous operator on S′(Rd)

• too. Note also that OpW (f) is self-adjoint on L2(Rd) when f is real valued.

The Weyl quantization on T2d is obtained by restricting OpW (f) to certain subspaces of S′(Rd) when f ∈ C∞(T2d) (i.e. is Z2d periodic). The construction is as follows (see [7] for more details). For any ξ = (ξq, ξp) ∈ R2d, the phase space translation operator U(ξ) is defined by U(ξ)ψ(q) = ψ(q − ξq) exp i

• ξp · q − ξq · ξp

2

• ,

ψ ∈ S(Rd) 3

and is clearly a unitary operator on L2(Rd). One easily checks that U(ξ) = OpW (χξ), χξ(q, p) = exp i (q · ξp − p · ξq), and that the following Weyl-Heisenberg relations hold for all ξ, η ∈ R2d U(ξ)U(η) = exp i 2ω(η, ξ)U(ξ + η), (2.3) with ω the symplectic form defined by ω(ξ, η) = ξq · ηp − ξp · ηq. This relation shows in particular that, if n, m ∈ Z2d, U(n) and U(m) commute if and only if there exists N ∈ N such that 2πN = 1. (2.4) Since U(ξ) acts naturally on S′(R), we can introduce for any κ ∈ [0, 2π)2d the space H(κ) = {ψ ∈ S′(Rd) | U(n)ψ = eiω(κ,n)+i

nq·np 2 ψ,

∀ n = (nq, np) ∈ Z2d} and it turns out that H(κ) is of dimension N d if (2.4) holds (0 otherwise) with the basis ψκ

r (q) = N − d

2

k∈Z2d

eiκp·kδ0

• q − k − r

N − κq 2πN

• ,

r ∈ {0, · · · , N − 1}d. The latter is proven in [7] as well as the existence of a unique scalar product on each H(κ) making the above basis orthonormal and U(n/N) unitary for all n ∈ Z2d. The Weyl quantization on T2d is then defined by f ∈ C∞(T2d) → OpW (f)|H(κ). This is indeed a mapping from C∞(T2d) to L(H(κ)), i.e. H(κ) is stable under OpW (f), since

• ne can easily check that for any Z2d periodic function f

OpW (f) =

• n∈Z2d

fnU(n/N) if f(x) =

n fne2iπω(x,n), x ∈ R2d. Let us emphasize that the spaces H(κ) are very natural in

view of the following direct integral decomposition [7] L2(Rd) ≃ (2π)−2d ⊕

[0,2π)2d H(κ) dκ

in which the operators OpW (f)|H(κ) are the fibers of OpW (f) for this decomposition. To streamline the discussion we will write both quantizations on R2d and T2d under a single

• form. From now on, M will denote either R2d or T2d. The Weyl quantization on M can then be

defined as the map f ∈ B(M) → OpW (f) ∈ L(H) where B(M) is either B(R2d) or C∞(T2d) and H is either L2(Rd) or H(κ) (we omit the , κ dependence in the notations). In order to write OpW (f) in a unified way, we need to introduce the symplectic Fourier transform F on M defined by Ff(ξ) =

• M

exp(iω(ξ, x))f(x) dx 4

where ξ = (ξq, ξp)1 belongs to M∗ = R2d if M = R2d and (2πZ)2d if M = T2d. Then the following inversion formula holds f(x) =

• M∗ exp iω(x, ξ)Ff(ξ) dν(ξ)

(2.5) with dν = (2π)−2d× the Lebesgue measure (resp.

Z2d δ2πn) if M∗ = R2d (resp. (2πZ)2d) and

the Weyl quantization can easily be seen to be OpW (f) =

• M∗ U(ξ)Ff(ξ) dν(ξ).

(2.6) Of course, when M = R2d, all the integrals must be understood in the weak sense (in (2.6) we use the fact that U(ξ)ψ1, ψ2 belongs to S(R2d

ξ ) if ψ1, ψ2 ∈ S(Rd)). Note also the existence of C, ¯

d such that, if ||.||∞ denotes the L∞ norm on M,

• OpW (f)
• H→H ≤ C sup

|γ|≤ ¯ d

||∂γf||∞, ∀ f ∈ B(M). (2.7) This comes from (2.2) if M = R2d and from the unitarity of U(n/N) combined with the elementary estimate

n |fn| ≤ C sup|γ|≤2d+1 ||∂γf||∞ when M = T2d.

This completes the definition of the Weyl quantization on M. Regarding the composition of the corresponding operators, we have the Proposition 2.1. There exists a bilinear map (f, g) → f#g from B(M)2 to B(M) such that OpW (f)OpW (g) = OpW (f#g). The function f#g has a full asymptotic expansion in powers of , meaning that for all integers J f#g =

• j<J

jf#jg + Jr

J(f, g)

where f#jg =

|α+β|=j Γ(α, β)∂α q ∂β p f∂β q ∂α p g, with Γ(α, β)−1 = (−1)αα!β!(2i)|α+β| and for all

γ ∈ N2d

• ∂γr

J(f, g)

• ∞ ≤ CJ+|γ|

d

J! sup

|γ1|≤J+|γ|+ ˜ d

||∂γ1f||∞ sup

|γ2|≤J+|γ|+ ˜ d

||∂γ2g||∞, 0 < ≤ 1. (2.8) for some constants Cd, ˜ d depending only on d. Note that (.#j.) is symmetric (resp. skew symmetric) for j even (resp. odd) and that g#1f − f#1g = −i(∇pg · ∇qf − ∇qg · ∇pf) = −i{g, f}. (2.9)

• Proof. This result is well known if M = R2d (see for instance the appendix of [8] for a simple

proof). We briefly sketch the proof in the case M = T2d. Using (2.3) and (2.6), we have OpW (f)OpW (g) =

• M∗
• M∗ eiω(η,ξ)/2U((ξ + η))Ff(ξ)Fg(η) dν(ξ)dν(η)

(2.10) =

• M∗ U(ξ)
• M∗ eiω(η−ξ,η)/2Ff(ξ − η)Fg(η) dν(η)
• dν(ξ). (2.11)

1Throughout this paper, x will denote the running point of M and ξ the one of M∗, unlike the usual notation

• f microlocal analysis where (x, ξ) is the running point of R2d.

5

Expanding eiω(η−ξ,η)/2 by the Taylor formula, we get the expansion of f#g with a remainder rJ = r

J(f, g) defined by its Fourier transform as follows

FrJ(ξ) = (−2i)−J (J − 1)! 1 (1 − t)J−1

• M∗ eitω(η−ξ,η)ω(η − ξ, η)JFf(ξ − η)Fg(η) dν(η) dt.

(2.12) Since |ξγFrJ(ξ)| = |F∂

γrJ(ξ)|, with

γ = (γp, γq) if γ = (γq, γp) (γq, γp ∈ Nd), we have to consider ξγω(η − ξ, η)J =

• γ1+γ2=γ
• |β|=J

(−1)|βq|+|γ1| γ! γ1!γ2! J! β!(η − ξ)β+γ1η

• β+γ2

where β = (βq, βp) with βq, βp ∈ Nd. The sum contains at most CJ+|γ|

d

terms and since J! β! ≤ (2d)J, γ! γ1!γ2! ≤ 2|γ| we conclude that (2.8) is now a simple consequence of the fact that

• M∗
• ξ

αFf(ξ)

• dν(ξ) =
• M∗ |F∂αf(ξ)| dν(ξ) ≤ Cd

sup

|α1|≤2d+1

||∂α1+αf||∞. We omit the details.

• Remark. The above proof can be repeated verbatim if M = R2d and B(R2d) is replaced by S(R2d).

We now present a unified version of Egorov Theorem, that is the semiclassical analysis of eitO

pW (g)/OpW (f)e−itO pW (g)/ for f, g ∈ B(M), with g real valued. This result is well known for

M = R2d [10, 22, 16, 18] and the purpose of what follows is essentially to prove a similar result for M = T2d, with an explicit remainder term. The result is based on the following simple remark: if A is a bounded self-adjoint operator and B(t) is a strongly C1 family of bounded operators, then eitA/B(0)e−itA/ − B(t) = i

• t

ei(t−s)A/

• i d

dsB(s) + [A, B(s)]

• e−i(t−s)A/.

(2.13) We shall use this formula with A = OpW (g) and B(t) of the form B(t) =

• j<J

jOpW (fj(t)) with f0(t), · · · , fJ−1(t) ∈ B(M) such that

j<J jfj(0) = f (i.e. B(0) = OpW (f)) and

i d dsB(s) + [A, B(s)] = O(J+1) (2.14) where O(hJ+1) is to be understood the operator norm on H. Expanding [A, B(s)] in powers of by means of Proposition 2.1, (2.14) leads to the following conditions on the functions fj(s) ∂sf0 − {g, f0} = 0, f0(0) = f, (2.15) ∂sfj − {g, fj} = 2i

• l+k=j+1

g#kfl, fj(0) = 0 for j ≥ 1 (2.16) 6

where, in the last sum, 3 ≤ k ≤ J − l is odd and l ≤ J − 1, which implies actually that l ≤ j − 2. This system is thus triangular and can be solved using the Hamiltonian flow φs of g, since the solution of ∂sa − {g, a} = b with as=0 = a0 is given by a(s, x) = a0(φs(x)) + s b(τ, φs−τ(x)) dτ, x ∈ M. Note that if M = T2d and g is identified with a Z2d periodic function on R2d, the associated Hamiltonian flow ˜ φs on R2d is easily seen to satisfy the identity ˜ φs(x + n) = ˜ φs(x) + n for all x ∈ R2d and n ∈ Z2d. This shows that the formulas for the fj(s) are the same for M = R2d and T2d, if f and g are Z2d periodic. Let us now define the linear operators Ls

j on B(M) by Ls jf := fj(s). We have

Ls

0f = f ◦ φs,

Ls

j ≡ 0 for j odd

(2.17) the latter being a consequence of the (skew) symmetry of #k for k (odd) even. For j ≥ 2 even, an induction shows that Ls

j =

1 (2i)j

j/2

• k=1
• m1+···+mk=j/2
• |α1+β1|=1+2m1

· · ·

• |αk+βk|=1+2mk

s · · · sk−1 Ls−s1 M α1,β1Ls1−s2 · · · M αk,βkLsk

0 dsk · · · ds1

(2.18) where m1 ≥ 1, · · · , mk ≥ 1 in the sum and M α,β is the differential operator M α,β = (−1)|α| α!β! ∂β

q ∂α p g∂α q ∂β p .

(2.19) Taking the remainders into account, one gets the following result: Theorem 2.2 (Egorov theorem). For all f, g ∈ B(M) with g real valued and all J ≥ 1 we have eitO

pW (g)/OpW (f)e−itO pW (g)/ =

• j<J

jOpW (Lt

jf) + JRt J(f, )

where the operator Rt

J(f, ) has the following explicit form

i t ei(t−s)O

pW (g)/

• l<J

OpW (r

J−l+1(g, Ls l f)) − OpW (r J−l+1(Ls l f, g))

• e−i(t−s)O

pW (g)/ ds.

Note that estimates on ||Rt

J(f, )||H→H can then be derived from (2.7), (2.8) and estimates on

the derivatives of Ls

• jf. This will be extensively used in the next section.

3 Perturbations of quantized hyperbolic maps

In this section, we address the problem of the semi-classical approximation of U −t

ǫ OpW (f)U t ǫ as

↓ 0 in the Ehrenfest time limit t ≈ | ln |, when Uǫ is a unitary operator on H of the form Uǫ = e−iǫO

pW (g)/M(A)

7

with M(A) the quantization of a symplectic matrix with integer entries A ∈ Sp(d, Z). We refer to [7, 4, 5] and [14] for the definition of M(A) by mean of the metaplectic representation of Sp(d, R) and only quote the properties that we need. The operator M(A) is defined, up to a phase, as the unique operator on S′(Rd) such that M(A)−1OpW (f)M(A) = OpW (f ◦ A), ∀ f ∈ B(R2d). (3.1) If M = R2d, M(A) is unitary on L2(Rd), but if M = T2d and H = H(κ) one has to choose special values of κ to ensure that M(A) maps H(κ) into itself, in which case M(A) is unitary (see [7] for more details); from now on, we shall assume that such a choice, which depends on , has been

• made. Then, (3.1) holds on M = R2d and T2d and this is often expressed by saying that for linear

evolutions ‘Egorov is exact’, meaning there is no remainder term. Let us now describe the results of this section. We will denote by φǫ the Hamiltonian flow associated to a fixed real valued g ∈ B(M) and consider the discrete group (Φt

ǫ)t∈Z of symplecto-

morphisms on M defined by Φǫ = φǫ ◦ A. (3.2) Then, by setting ˜ Ljf = (Lǫ

jf) ◦ A

with the notations of (2.17) and (2.18), we can consider the functions Lt

0f = f ◦ Φt ǫ,

Lt

jf =

• l1+···+lt=j

˜ Ll1 · · · ˜ Lltf defined for j ≥ 1, t ≥ 0 integers and f ∈ B(M). Note that they depend on ǫ but we omit this dependence for notational convenience. Note also that Lt

0 = (˜

L0)t and that Lt

j ≡ 0 if j is odd.

Our goal is to show that U −t

ǫ OpW (f)U t ǫ ∼

• hjOpW (Lt

jf),

↓ 0, in a scale of times t described in terms of exponents ΓA, Γg that we now define. For the sake of simplicity, we shall assume that A is diagonalizable over R, meaning that there exists an invertible matrix P with real entries such that A = P −1DP with D diagonal. Note that such a condition is of course satisfied if A is symmetric, e.g. the cat map. At the end of the section, we explain how to cope with general symplectic matrices A ∈ Sp(d, Z). Let us define ΓA ≥ 0 by eΓA = sup

σ(A)

|λ|. Of course, this quantity is well defined for any invertible matrix A with real or complex spectrum. For z = (z1, · · · , z2d) ∈ C2d, we denote by |z| := (|z1|2 + · · · |z2d|2)1/2 its standard hermitian norm and set ||z||P := |Pz|. The interest of the norm ||.||P is that we have ||Az||P ≤ eΓA||z||P , ||Im Az||P ≤ eΓA||Im z||P ∀ z ∈ C2d, (3.3) which we shall use extensively in the sequel. Then, inspired by [23, 8], we define the open sets Ωδ ⊂ C2d for δ > 0 by Ωδ = {z ∈ C2d | ||Im z||P < δ} 8

and we consider the family of norms ||.||τ,δ defined for τ ∈ (0, 1) by ||f||τ,δ = sup

z∈Ωτδ

|f(z)| for functions f which are bounded and analytic on Ωδ. We can now set Γg = sup

z∈Ωδ

|||J ∇2g(z)|||P , J = I −I

• with ∇2g the Hessian matrix of g and |||B|||P := supz=0 ||Bz||P /||z||P for B ∈ M2d(C). Note that

Γg = 0 unless g is constant which is a trivial situation. We then define Γǫ = ΓA + ǫΓg and our main result is the following: Theorem 3.1. Assume that f, g ∈ B(M), with g real valued, have bounded and analytic extensions to Ωδ for some δ > 0. Then, for all 0 < ν < 2, there exists J0 > 0 such that for all J > J0 U −t

ǫ OpW (f)U t ǫ =

• j<J

jOpW Lt

jf

• + J̺t

J(f, ǫ, )

(3.4) with a remainder such that, for all 0 ≤ ǫ ≤ 1,

• J̺t

J(f, ǫ, )

• H→H → 0

as → 0 if 0 ≤ t ≤ 2 − ν 3Γǫ | ln |. (3.5) The reader may wonder what (3.5) means if Γǫ = 0. In such a case Γg = 0 thus g is constant so (3.4) becomes U −t

ǫ OpW (f)U t ǫ = OpW (f ◦ At) by (3.1) which holds for all t ≥ 0. In Section 3, we

will anyway be interested in the situation where ΓA > 0 and ǫ is small so that Γǫ > 0. We also emphasize that the analyticity assumption is imposed by our need to control high order derivatives of f and g in order to estimate ̺t

J(f, ǫ, ). Similarly to [8], we could probably relax such

a condition by considering quasi-analytic functions (e.g. Gevrey functions) which would allow us to consider compactly supported f. The rest of this section is now devoted to the proof of Theorem 3.1. The principle is rather simple and is the following: a straightforward application of Theorem 2.2 shows that U −1

ǫ

OpW (f)Uǫ =

• j<J

jOpW (˜ Ljf) + JM(A)−1Rǫ

J(f, )M(A),

∀ J > 0, hence an induction on t ≥ 1 shows that (3.4) holds with ̺t

J(f, ǫ, ) = tJ

• j=J

j−JOpW (Kt

j,Jf) + t

• s=1

U s−t

ǫ

M(A)−1Rǫ

J(Es−1 J

f, )M(A)U t−s

ǫ

(3.6) with the operators Kt

j,J and Et J defined by

Kt

j,J =

• l1+···+lt=j

l1<J,··· ,lt<J

˜ Llt · · · ˜ Ll1, Et

J =

• l1<J

· · ·

• lt<J

l1+···+lt ˜ Llt · · · ˜ Ll1. (3.7) Note that Kt

j,J depends on both j and J unless j < J in which case Kt j,J = Lt

• j. Note moreover

that Et

J depends on and that we set K1 j,J = 0, E0 J = id.

Thus (2.7) reduces the proof of Theorem 3.1 essentially to estimate the derivatives of ˜ Llt · · · ˜ Ll1f. To that end, we shall use the following extension of a lemma of [8]. 9

Lemma 3.2. There exists a constant CP depending only on P such that, if ||f||τ,δ ≤ M CP 1 − τ a , ∀ 0 < τ < 1 (3.8) for some M, a ≥ 0, then for all γ we have ||∂γf||τ,δ ≤ M(a + |γ|) · · · (a + 1)δ−|γ| CP 1 − τ a+|γ| ∀ 0 < τ < 1.

• Proof. In [8], the authors show that the result holds with P = I and CP = e. Our lemma follows

from their result applied to f ◦ P −1.

• In order to estimate f ◦ Φǫ we will capitalize on two facts: on one hand, (3.3) implies that

||f ◦ A||τ,e−ΓAδ ≤ ||f||τ,δ (3.9) and on the other hand we have, for any 0 ≤ s ≤ t, ||Im z||P ≤ τδe−tΓg ⇒ ||Im φt−s(z)||P ≤ τδe−sΓg. The latter is actually shown in [8] only for P = I but the very same method easily leads to this

• estimate. We therefore omit the proof and rather emphasize that it implies that

||f ◦ φt||τ,δe−tΓg ≤ ||f||τ,δ (3.10) which leads to the Lemma 3.3. Assume that Cg > 0 is such that |∂γg(z)| ≤ γ!C|γ|

g

for all z ∈ Ωδ and all |γ| ≥ 1. Assume moreover that (3.8) holds. Then for all j ≥ 2 even and all s ≥ 0 real, we have ||Ls

jf||τ,δe−sΓg ≤ M

CP 1 − τ a+ 3j

2

e

3j 2 sΓg(a + 1) · · ·

• a + 3j

2

• (4dCg/δ)3j/2

j/2

• k=1

sk k! , for all τ ∈ (0, 1).

• Proof. We first note that, by an easy induction on k ≥ 0, the following result hods: if s0, · · · , sk

are non negative real numbers such that s0 + · · · + sk = s and α1, β1, · · · , αk, βk are non zero multi-indices such that |α1 + β1| + · · · + |αk + βk| = n, then for all τ ∈ (0, 1) ||Lsk

0 M αk,βk · · · Ls1 0 M α1,β1Ls0 0 f||τ,δe−sΓg ≤ M

CP 1 − τ a+n Cn

g δ−nesnΓg(a + 1) · · · (a + n).

This follows from Lemma 3.2 and (3.10) (recall that M α,β is defined by (2.19)). The lemma is then a consequence of (2.18) combined with the above estimate, the fact that #{(α, β) ∈ N2d | |α + β| = 1 + 2m} = (2d + 2m)! (2d − 1)!(2m + 1)! ≤ (2d)1+2m, #{(m1, · · · , mk) ∈ Nk | m1 + · · · + mk = j/2} = (k − 1 + j/2)! (k − 1)!(j/2)! ≤ 2k−1+ j

2 ,

and the fact that s

0 · · ·

sk−1 dsk · · · ds1 = sk/k!.

• We can now state the main ingredient of the proof of Theorem 3.1.

10

Proposition 3.4. With the same assumptions as in Lemma 3.3, we have: for all j ≥ 2 and all integers l1, · · · , lt such that l1 + · · · + lt = j, we have for all ǫ ∈ [0, 1] ||˜ Llt · · · ˜ Ll1f||τ,δe−tΓǫ ≤ M CP 1 − τ a+ 3j

2

e

3t 2 jΓǫ(a + 1) · · ·

• a + 3j

2

• (4dCgeǫ/δ)3j/2

provided that (3.8) holds. In addition, if |f(z)| ≤ M on Ωδ, there exists a constant K such that, for all t ≥ 1, all γ and all ǫ ∈ [0, 1] ||∂γKt

j,Jf||∞ ≤ MtjK(1+ǫ)j(|γ| + 3j/2)!etΓǫ(|γ|+3j/2),

0 ≤ j ≤ tJ. (3.11)

• Proof. Recall that we can assume that j is even. We obtain the first statement by induction on t ≥ 1

using lemma 3.3 with s = ǫ and (3.9) which we use through ||f ◦ A||τ,e−3ΓA/2δ ≤ ||f ◦ A||τ,e−ΓAδ ≤ ||f||τ,δ. This, together with (3.7) then yields the second statement since #{l1 + · · · + lt = j} ≤ tj.

• Proof of Theorem 3.1. We first estimate || tJ

j=J hjOpw(Kt j,Jf)||H→H. Using (2.7), (3.11) allows

to estimate ln

• ||hjOpw(Kt

j,Jf)||H→H

• from above by

j

• ln + t

3 2 + ¯ d J

• Γǫ + ln(K2t) +

3 2 + ¯ d J

• ln( ¯

d + 3j/2)

• + ln CM

using the fact that 1/j ≤ 1/J and that ǫ ∈ [0, 1]. By choosing J large enough we can assume that 3 2 + ¯ d J ≤ 3 2 − ν/2. Since j ≤ Jt, the term ln( ¯ d + 3j/2) is O(ln | ln |||) as ↓ 0 thus we get the existence of a new constant C such that for all ν ∈ (0, 2), ∈ (0, 1], ǫ ∈ [0, 1], 1 ≤ t ≤ (2 − ν)/3Γǫ and j ∈ [J, tJ] ||hjOpw(Kt

j,Jf)||H→H ≤ Cνj/2 (ln(C + | ln |))Cj ≤ ˜

Cνj/4. (3.12) Since

J≤j≤tJ contains O(| ln |) terms, we see that || tJ j=J hjOpw(Kt j,Jf)||H→H → 0.

Now the norm of second term of (3.6) multiplied by J can be estimated by tJhJ sup

0≤τ≤t−1 s∈[0,ǫ], l<J

||OpW (r

J−l+1(g, Ls l Eτ Jf))||H→H + ||OpW (r J+l−1(Ls l Eτ Jf, g))||H→H

(3.13) with the notations of Theorem 2.2. We proceed as before to estimate Ls

l Eτ Jf and we obtain the

theorem.

• Let us now briefly describe how to prove such results for a general A ∈ Sp(d, Z) with ΓA > 0.

We claim that, in this case, we have the following result: for any ˜ ΓA > ΓA there exists an invertible matrix P with real entries such that |||P −1AP||| ≤ e

˜ ΓA

(3.14) where |||.||| is the matrix norm associated to the hermitian norm |.| on R2. We can prove this statement as follows. Assume first that the spectrum of A is real and let us choose a basis 11

(e1, · · · , e2d) of R2d in which A is in Jordan normal form. If (ej, · · · , ej+p) corresponds to a Jordan block J(λ) =          λ 1 · · · λ 1 ... . . . . . . ... ... ... . . . ... ... 1 · · · · · · λ          , then by changing (ej, ej+1, · · · , ej+p) into (ej, εej+1, · · · , εpej+p) with ε > 0, the above block is changed into the same one with 1 replaced by ε. Proceeding similarly for all the blocks, we

• btain the existence of basis in which A is the sum of a diagonal matrix of norm eΓA and of a

nilpotent matrix of norm O(ε). This leads to the statement when the spectrum is real. For non real eigenvalues λ = ρeiθ, using Jordan normal form over C2d, we have to consider blocks of the form J(λ) J(¯ λ)

• .

It is then standard that there exists a basis of real vectors in which the endomorphism represented by the above block has a matrix of the form N + ρR(θ) where N is nilpotent and R(θ) is block diagonal matrix of rotations (of dimension 2) of angle θ. Then, by changing this basis as in the case of a real spectrum, we can assume that N is small and we obtain (3.14) in the general case.

4 Equirepartition of time-evolved localized states

4.1 The example of (generalized) coherent states

In this subsection, we shall prove that the generalized coherent states, defined below, when evolved

• ver sufficiently long times, equidistribute on the torus.

To define the states in question, we proceed as follows. Let ϕ(q) = h−µ/2ϕ q µ

• (4.1)

with ϕ ∈ S(Rd),

• |ϕ|2 = 1 and µ ∈ (0, 1). Then we set

ϕa

= U(a)ϕ

(4.2) which defines a family of states in L2(Rd) indexed by a ∈ R2d. These are commonly referred to as (generalized) coherent states. The corresponding states on the torus, i.e. belonging to H(κ), are defined by ϕa

,κ := S(κ)ϕa =

 

np∈Zd

e−iκq·npU(0, np)    

nq∈Zd

eiκp·nqU(nq, 0)   ϕa

• (4.3)

which converges in S′(Rd) (see [7]). The main property of these states that we shall use is

• ϕa

,κ, OpW (f)ϕb ,κ

• H(κ) =
• n∈Z2d

(−1)Nnq·npeiω(κ,n)eiω(n,b)/2 ϕa

, OpW (f)ϕb−n

• L2(Rd)

(4.4) 12

which is proven in [4]. The best known example of such functions are obtained by choosing µ = 1/2 and ϕ(q) = η(q) := π−d/4e−q2/2. With this choice one obtains the standard coherent states. If ˜ ϕ is another Schwartz function and ˜ ϕ is defined similarly to (4.1), the Wigner function W(x) associated to ϕ, ˜ ϕ is defined by

• ϕ, OpW (f) ˜

ϕ

• L2(Rd) =
• R2d f(x)W(x) dx

for all f ∈ B(R2d). For general ϕ, ˜ ϕ in L2(Rd), W is a distribution, but for Schwartz functions it is a Schwartz function as well given by W(x) = (2π)−d

• e−i˜

q·p/ϕ(q − ˜

q/2) ˜ ϕ(q + ˜ q/2) d˜ q, x = (q, p). With the simple dependence considered in (4.1), it is easy to see that the Wigner function W (a,b)

• (x) associated to U(a)ϕ and U(b) ˜

ϕ takes the following form for any a, b ∈ R2d W (a,b)

• (x) = e−iω(a,b)/2+iω(x,b−a)/−dW1
• Σµ
• x − a + b

2

• (4.5)

where Σµ

is the linear map on R2d defined by Σµ (q, p) = (q/µ, p/1−µ) and W1 the Wigner

function of ϕ, ˜ ϕ. Note that, since W1 ∈ L1(R2d), (4.5) implies that ||W (a,b)

• ||L1 = ||W1||L1 is

independent of . Note also that when ϕ(q) = ˜ ϕ(q) = η(q), one easily checks that W1(x) = π−de−x2 (4.6) which makes (4.5) completely explicit in this case. Our main result is Theorem 4.7. As explained in the introduction, its proof goes in two steps. First we use the Egorov theorem to establish that on a suitable time scale

• U t

ǫϕa ,κ, OpW (f)U t ǫϕa ,κ

• H(κ)

is equivalent to

• ϕa

, OpW (f ◦ Φt ǫ)ϕa

• L2(Rd) (Proposition 4.2). Then we use an estimate on the clas-

sical evolution (exponential mixing) to control this last term. As a warm up for the first step, we show for a particularly simple class of states how the Egorov expansion (3.4) can be reduced to the first term. Proposition 4.1. Let Ψ ∈ H be a family such that there exists C satisfying

• Ψ, OpW (f)Ψ
• ≤ C||f||∞,

0 < ≤ 1 (4.7) for all f in B(M) having a bounded and analytic continuation to some Ωδ. Then

• U t

ǫΨ, OpW (f)U t ǫΨ

• H −
• Ψ, OpW (f ◦ Φt

ǫ)Ψ

• H → 0,

→ 0 provided 0 ≤ t ≤ (2 − ν)| ln |/3Γǫ for some ν ∈ (0, 2).

• Proof. Using Theorem 3.1, we only have to show that for all 1 ≤ j < J we have

j Ψ, OpW (Lt

jf)Ψ

• H → 0,

→ 0 in the specified range of times. This readily follows from the fact that j||Lt

jf||∞ ≤ Cjj| ln |je−j(1− ν

2) ln

13

by estimate (3.11), where one should recall that Lt

j = Kt j,J if j < J.

• The condition (4.7) is for instance satisfied by coherent states, in both cases M = R2d and
• T2d. This readily follows from the independence of ||W (a,a)
• ||L1 if M = R2d. In case of the torus,

it is a simple exercise using the Poisson summation formula. Note also that, if f is periodic (in particular if M = T2d), we can get rid of the analyticity of f since it is the uniform limit of a sequence of trigonometric polynomials. Nevertheless, regarding coherent states on the torus, the above result is not precise enough for

• ur purpose since the term
• ϕa

,κ, OpW (f ◦ Φt ǫ)ϕa ,κ

• is not very explicit. This is why we give the

next proposition whose proof will also be used in the proof of Theorem 4.9. Proposition 4.2. Fix a ∈ R2d and assume that 0 < µ < 1. Then, for all f ∈ C∞(T2d), we have

• U t

ǫϕa ,κ, OpW (f)U t ǫϕa ,κ

• H(κ) −
• ϕa

, OpW (f ◦ Φt ǫ)ϕa

• L2(Rd) → 0,

→ 0 provided 0 ≤ t ≤ (2 − ν)| ln |/3Γǫ for some ν ∈ (0, 2).

• Proof. Let us first note that, by truncating the Fourier series of f, there exists a sequence fM of Z2d

periodic analytic functions such that fM → f in B(T2d). Since ||OpW (f) − OpW (fM)||H→H → 0 and

• ϕa

, OpW (f ◦ Φt ǫ)ϕa

• L2(Rd) −
• ϕa

, OpW (fM ◦ Φt ǫ)ϕa

• L2(Rd) → 0,

M → +∞ uniformly with respect to t ∈ R and ∈ (0, 1] by (4.5), we are left with the case where f is analytic. Then, by Theorem 3.1, we only have to study the difference

• j<J

j ϕa

,κ, OpW (Lt jf)ϕa ,κ

• H(κ) −
• ϕa

, OpW (f ◦ Φt ǫ)ϕa

• L2(Rd) ,

thus the result will follow from (4.4) if we show that, in the specified range of times,

• n=0
• ϕa

, OpW (f ◦ Φt ǫ)ϕa−n

• L2(Rd)
• →

0, (4.8) j

n∈Z2d

• ϕa

, OpW (Lt jf)ϕa−n

• L2(Rd)
• →

j ≥ 1. (4.9) We first note that the term corresponding to n = 0 in (4.9) has been studied in the proof of the previous proposition, and its limit is 0. We may therefore assume that n = 0 in both sums. Using (4.5), integrations by parts with 2∆x/|n|2 show that, for all j ≥ 0 and all M > 0

• j

ϕa

, Opw(Lt jf)ϕa−n

• L2(Rd)
• ≤ Cj|n|−2M
• |γ|≤2M

j+2M−m(2M−|γ|)||∂γLt

jf||∞,

where m = max(µ, 1 − µ). We get the result by the simple observation that j+2M−m(2M−|γ|)||∂γLt

jf||∞ ≤ Cf2νj+CM → 0,

→ 0 for |γ| ≤ 2M, with C = 2(1 + ν)/3 if m < (2 − ν)/3 and C = 2(1 − m) otherwise. This follows from (3.11) by distinguishing both cases m ≥ (2 − ν)/3 and m < (2 − ν)/3.

• This proposition, combined with (4.5) allows us to reduce the study of the matrix elements of

evolved coherent states to a problem in classical dynamics. By this, we mean that the main result 14

• f this section, Theorem 4.7, is a direct consequence of Proposition 4.2 and of the mixing estimates

given in the Appendix A. Note that from now on, we shall be working with d = 1. As explained in the introduction, the reason for this is that, whereas the mixing rate is controlled by the smallest Lyapounov exponent

• f A, the error in the Egorov theorem is controlled by its largest Lyapounov exponent.

As a warm-up, and in order to bring out the main strategy, we first prove a simplified version

• f the result:

Theorem 4.3. Assume that ΓA > 0. Let a in R2, f ∈ B(T2) and 1/3 < µ < 2/3. Then, for all ν > 0 there exists ǫ0 small enough (independent of f) such that for |ǫ| < ǫ0 we have

• U t

ǫϕa ,κ, OpW (f)U t ǫϕa ,κ

• H(κ) →
• T2 f(x) dx,

→ 0, provided that m + ν Γǫ | ln | ≤ t ≤ 2 − ν 3Γǫ | ln |, m = max(µ, 1 − µ). (4.10)

• Proof. We first remark that, by choosing 0 < Γ < ΓA close enough to ΓA and ǫ small enough we

have 1 > Γ Γǫ > 1 + ν/2 1 + ν . (4.11) Combined with (4.10), this estimate implies that t/| ln | > (m + ν/2)/Γ and thus e−tΓ ≤ m+ ν

2 .

(4.12) By Proposition 4.2 and (4.5) we only have to study the limit of

• R2(f ◦ Φt

ǫ)(x)W (a,a)

• (x) dx

(4.13) for which

• W (a,a)
• (x)dx = 1. Choosing a smooth cutoff function χ so that χ = 1 near 0 and which

is supported close to 0, then setting g(x) := W (a,a)

• (x)χ(x−a), we have ||W (a,a)
• −g||L1 = O(h∞)

thus

• R2(f ◦ Φt

ǫ)(x)W (a,a)

• (x) dx −
• R2(f ◦ Φt

ǫ)(x)g(x) dx → 0,

↓ 0 uniformly with respect to t ∈ R. The last integral can obviously be interpreted as an integral over T2 since g is supported close to a and consequently we can use Corollary A.2. The result now simply follows from the fact that e−tΓ||g||W 1,1 = O(h−m)e−tΓ → 0 by (4.12).

• The above proof is a rather direct application of Proposition 4.2 and Corollary A.2 but it fails

if m ≥ 2/3 (i.e. µ / ∈ (1/3, 2/3)) since e−tΓh−m > 1, for, in that case, e−tΓ > 2/3. The problem stems from the lower bound in (4.10), which arises because ∇W (a,a)

• behaves like −m.

One expects on intuitive grounds that it should be possible to replace m by m = min(µ, 1 − µ) which is

• f course less than 1/2 which is less than 2/3. We shall prove this is true, but for that purpose we

will need to exploit some more detailed knowledge about the Anosov diffeomorphisms we study. The trick consists in applying a well known idea in the theory of Anosov systems: it is possible to replace (4.13) by an expression obtained by performing an integral along the stable foliation. 15

Since the evolution stretches the function W (a,a)

• along the unstable manifold, this corresponds to

smoothening out the fastest oscillations in W (a,a)

• , replacing the latter by a function that has a

derivative controlled by −m. Let us start the proof. By Proposition 4.2, we have to study (4.13) where W (a,a)

• can be replaced, as in the proof of Theorem 4.3, by g which we can assume to be

supported as close to a as we want. This will allow us to use the following result. Theorem 4.4. [3, 2] For all Γ < ΓA, there exists ǫ0 small enough such that for all |ǫ| < ǫ0 the following holds: there exist σǫ > 0 and a C1+σǫ diffeomorphism x → Fǫ(x) = (s(x), u(x)), from a neighborhood of a ∈ T2 to a neighborhood of 0 ∈ R2 such that Fǫ(a) = 0 and

• ∂s
• f ◦ Φt

ǫ ◦ F −1 ǫ

• (u, s)
• ≤ Cfe−Γt

(4.14) for all t ≥ 0, all (u, s) in the neighborhood of 0 and all f ∈ C1(T2, R). Here C1+σ denotes the corresponding H¨

• lder class.

Using this result, we can perform the following change of variables

• R2
• f ◦ Φt

ǫ

• (x)g(x) dx =

f ◦ Φt

ǫ ◦ F −1 ǫ

• (u, s)
• g ◦ F −1

ǫ

• (u, s)Jǫ(u, s) duds

(4.15) where Jǫ ∈ Cσǫ. On the right hand side of this equation, we eventually want to use Corollary A.2, but the Cσǫ regularity of Jǫ(u, s) is not sufficient for that purpose. Fortunately, the term Jǫ is essentially irrelevant in view of the following result. Lemma 4.5. i) (g ◦ F −1

ǫ

)0<≤1 is a bounded family in L1(R2). ii) For all θ ∈ (0, 1) there exists a family J

ǫ such that, if ||.||∞ is the sup norm over a fixed small

neighborhood of 0, ||J

ǫ − Jǫ||∞ ≤ Cσǫθ,

||∇J

ǫ ||∞ ≤ C−θ.

Proof. i) follows from (4.5) and ii) from a standard convolution argument by a C∞ function χ(u, s) = −2θχ(u/θ, s/θ).

• Using this lemma and (4.14), the right hand side of (4.15) takes the form

f ◦ Φt

ǫ ◦ F −1 ǫ

• (u, 0)k(u) du + O1(e−Γt) + O2(σǫθ)

where O1 is uniform with respect to ∈ (0, 1], O2 is uniform with respect to t ≥ 0 and where the C1 function k(u) is given by k(u) = g ◦ F −1

ǫ

• (u, s)J

ǫ (u, s) ds.

Note that k is bounded in L1 and that

• R k(u) du → 1 as → 0. The key remark is now that

the derivative of this function is essentially controlled by −m rather than by −m, as a rough estimate would show. That is the content of the following proposition. Note that, in what follows, q, p are the canonical coordinates of R2. They also define local coordinates on T2 close to any a, and this makes the following statement clear. Proposition 4.6. Assume that µ ≤ 1/2 (i.e. that max(µ, 1 − µ) = 1 − µ). Then, if the support of g is sufficiently close to a and if ∂s

• p ◦ F −1

ǫ

• (0, 0) = 0

(4.16) 16

then, there exists ˜ k ∈ C1(R) such that

• k − ˜

k

• L1 → 0 as → 0 and
• d˜

k/du

• L1 ≤ C−µ.

(4.17) Proof. The condition (4.16) shows that, if δ1, δ2 are small enough, s → (p ◦ F −1

ǫ

)(u, s) is a diffeomorphism from (−δ1, δ1) onto its range for each u ∈ (−δ2, δ2). Thus, if the support of g is small enough, we can use (p ◦ F −1

ǫ

)(u, s) as a new variable in the integral defining k so that it becomes k(u) = −1

• W1

˜ q(u, p) − q(a) µ , p − p(a) 1−µ

• χ

ǫ (u, p) jǫ(u, p)dp

with jǫ(u, p) the Cσǫ jacobian of the change of variable and χ

ǫ (u, p) the term corresponding to

χ(F −1

ǫ

(u, s) − a)J

ǫ (u, s). Changing again the variable with ˜

p = (p − p(a))/1−µ, we would get the result if jǫ was C1, by choosing θ = µ. We can overcome the non smoothness of jǫ by the same principle as for Lemma 4.5: we choose j

ǫ approaching jǫ uniformly on the support of

χ

ǫ , such that

∇j

ǫ = O(−µ) and then

˜ k(u) = −1

• W1

˜ q(u, p) − q(a) µ , p − p(a) 1−µ

• χ

ǫ (u, p) j ǫ (u, p)dp

has the expected properties.

• Remark. The condition (4.16) expresses the fact that, at the point a, the submanifold {q =

q(a)} is not aligned with the unstable manifold. Of course, if µ > 1/2, the same result holds if ∂s(q ◦ F −1

ǫ

)(0, 0) = 0. We are now ready for the proof of the main theorem of this subsection. Theorem 4.7. Assume that 0 < µ ≤ 1/3 (resp. 2/3 ≤ µ < 1) and that the unstable manifold through a is not aligned with the submanifold {q = q(a)} (resp. {p = p(a)}). Assume moreover that ΓA > 0. Then, there exists ǫ0 such that, for |ǫ| < ǫ0 and all f ∈ C∞(T2)

• U t

ǫϕa ,κ, OpW (f)U t ǫϕa ,κ

• H(κ) →
• T2 f(x) dx,

→ 0, provided that m + ν Γǫ | ln | ≤ t ≤ 2 − ν 3Γǫ | ln |, m = min(µ, 1 − µ). (4.18)

• Proof. The above discussion shows that we only have to prove that

f ◦ Φt

ǫ ◦ F −1 ǫ

• (u, 0)˜

k(u) du →

• T2 f(x) dx.

(4.19) Pick a smooth function ̺(s) supported close to 0 such that

• ̺(s)ds = 1. Then, using Theorem

4.4, the left hand side of (4.19) takes the form f ◦ Φt

ǫ ◦ F −1 ǫ

• (u, s)˜

k(u)̺(s) duds + O(e−Γǫt) 17

with O(e−Γǫt) uniform with respect to ∈ (0, 1]. This last integral is nothing but

• T2 f ◦ Φt

ǫ(x)˜

g(x) dx where ˜ g ◦ F −1

ǫ

(u, s) = ˜ k(u)̺(s)/Jǫ(u, s). Thus ˜ g is of the form ˜ g(1)

˜

g(2) with ˜ g(2) ∈ Cσǫ inde- pendent of and ||˜ g(1)

||L1 + m||∇˜

g(1)

||L1 = O(1). Note also that

• T2 ˜

g → 1 as → 0. Using Lemma 4.5 again to approach ˜ g(2) by C1 functions, we may assume that ˜ g is C1 and satisfies the same bound as ˜ g(1)

. We can now repeat the arguments of Theorem 4.3 and the result follows.

• 4.2

Semiclassical behavior of eigenstates

We now come to a more general result having applications in the description of the eigenvectors

• f Uǫ. Assume that Ψ,κ ∈ H(κ) satisfies, for all f ∈ C∞(T2d),
• Ψ,κ, OpW (f)Ψ,κ
• H(κ) → f(0),

↓ 0. (4.20) The right hand side could of course be replaced by f(a) for some a ∈ T2d or more generally by

• 1≤j≤J αjf(aJ) for finitely many points a1, · · · , aJ. Rather vaguely, this condition says that Ψ,κ

is concentrated at 0. This is confirmed by the following Lemma 4.8. There exists a sequence of positive numbers r → 0 and a family of functions χ ∈ C∞(T2d) supported in a ball of radius r centered at 0 (in T2d) such that 0 ≤ χ ≤ 1 and

• Ψ,κ − (2π)−d
• T2d χ(a)
• ηa

,κ, Ψ,κ

• H(κ) ηa

,κ da

• H(κ)

→ 0, ↓ 0. (4.21) Conversely, if (4.21) holds and ||Ψ,κ||H(κ) → 1 then (4.20) holds for all f ∈ C∞(T2d). The proof of this lemma is given in Appendix B, where we also recall basic results on coherent states decomposition over L2(Rd) and H(κ). Recall that ηa

,κ is defined by (4.1), (4.2) and (4.3)

with µ = 1/2 and η(q) = π−d/4e−q2/2. In the same way, we can define the concentration on a finite collection of points in a r neigh- borhood of those points. To simplify the notation, we set λ(a) = χ(a)

• ηa

,κ, Ψ,κ

• H(κ). The above lemma proves that

ψ,κ := (2π)−d

• T2d λ(a)ηa

,κ da

satisfies (4.20) as well and that

• Ψ,κ, U −t

ǫ OpW (f)U t ǫΨ,κ

• H(κ) −
• ψ,κ, U −t

ǫ OpW (f)U t ǫψ,κ

• H(κ) → 0,

↓ 0 uniformly with respect to t ≥ 0. This is the first step of the proof of the next theorem, in which the notations ., . and ||.|| stand for ., .H(κ) and ||.||H(κ) respectively. Theorem 4.9. Assume that ||Ψ,κ|| → 1 and that (4.21) holds for some sequence r such that r ≤ 1/2−σ, 18

with σ > 0. Then, as → 0,

• Ψ,κ, U −t

ǫ OpW (f)U t ǫΨ,κ

• − (2π)−2d
• T2d
• T2d λ(a)λ(b)
• ηa

,κ, OpW (f ◦ Φt ǫ)ηb ,κ

• dadb → 0

provided 0 ≤ Γǫt ≤ 1 2 + τ

• | ln |,

1 2 − 3τ − 4dσ > 0 and τ < 1 6. (4.22) If moreover d = 1, ΓA > 0 and τ − 5σ > 0, then there exists t → ∞ and ǫ(σ, τ) > 0 such that for all |ǫ| ≤ ǫ(σ, τ)

• Ψ,κ, U −t

ǫ

OpW (f)U t

ǫ Ψ,κ

• →
• T2 f(x) dx,

↓ 0. (4.23) This theorem generalizes a result of [6], Section 5, where only the case ǫ = 0 is treated. The proof is then much simpler, since there is then no error term in the Egorov theorem. The theorem says that, if a sequence of states concentrates sufficiently fast on a point a in T2, then the time evolved states equidistribute on the torus on some logarithmic time scale. Before proving this theorem, we show how it leads to a result on the semiclassical behaviour of the eigenvectors of Uǫ. Corollary 4.10. Assume that d = 1 and that ΓA > 0. For any 0 < σ < 1/38, there exists ǫ(σ) > 0 such that for all |ǫ| < ǫ(σ), no family Ψ,κ of eigenvectors of Uǫ can satisfy simultaneously (4.20) for all f and (4.21) with r ≤ 1/2−σ. We note in passing that a similar result (with a worse value of σ) holds for d > 1 provided we impose a pinching condition on the Lyapounov exponents of A as mentionned in the introduction. Roughly speaking, this corollary shows that, if a family of eigenvectors of Uǫ concentrates, in the semiclassical limit on a single point in phase space, then it must do so slowly. In other words, no such sequence can ’live’ in a ball of too small a radius r. In view of the comment after Lemma (4.8), it is clear that this result holds for a pure point measure supported on a finite number of periodic orbits. Given Theorem 4.9, the proof is very simple and identical to the case ǫ = 0 treated in [6], Section 5. We repeat it for completeness.

• Proof. For any 0 < σ < 1/38, one can find τ > 5σ satisfying (4.22). Furthermore, since Ψ,κ is an

eigenfunction,

• Ψ,κ, U −t

ǫ OpW (f)U t ǫΨ,κ

• =
• Ψ,κ, OpW (f)Ψ,κ
• for all t, thus by choosing t = t

and letting ↓ 0 we obtain f(0) =

• T2 f(x) dx

for all f ∈ C∞(T2d), which leads to a contradiction.

• Proof of Theorem 4.9. Here again, it is sufficient to assume that f is analytic. Using Theorem

3.1 and Lemma 4.8, it is clear that, if τ < 1/6 and tΓǫ ≤ (1/2 + τ)| ln |, we have

• Ψ,κ, U −t

ǫ OpW (f)U t ǫΨ,κ

• −
• j<J

j ψ,κ, OpW (Lt

jf)ψ,κ

• → 0,

↓ 0. The first part of the theorem will thus be proven if we show that, for any j ≥ 2 (recall that Lt

j ≡ 0

if j is odd), we have (2π)−2dj

• T2d
• T2d λ(a)λ(b)
• ηa

,κ, OpW (Lt jf)ηb ,κ

• dadb → 0,

↓ 0 19

if 1/2 − 3τ − 4dσ > 0. Using (4.4) and (4.5), integrations by parts similar to those of proposition 4.2 show easily that, for all M > 0,

• ηa

,κ, OpW (Lt jf)ηb ,κ

• =
• |n|≤C

(−1)Nnq·npeiω(κ,n)+iω(n,b)/2 ηa

, OpW (Lt jf)ηb−n

• L2 + O(M)

uniformly with respect to a, b ∈ [0, 1)2d and Γǫt ≤ (1/2 + τ)| ln |, with τ < 1/6. The constant C involved in the sum is such that |b − n − a| ≥ C−1|n| for all a, b ∈ [0, 1)2d and |n| > C. On the

• ther hand, using (4.5) and (3.11), one sees that, for any n ∈ Z2d and any j ≥ 2,

(2π)−2d

• T2d
• T2d |λ(a)λ(b)| j
• ηa

, OpW (Lt jf)ηb−n

• dadb ≤ C−2dr4d

1/2−3τ

since |λ(a)| ≤ ||Ψ,κ|| is bounded and λ is supported in a set of volume O(r2d

). The first part

• f the theorem follows.

We now prove the second part. Since χ can be chosen of the form χ(a) =

n1∈Z2d χ

• a+n1

r

• (see the Appendix B), it turns out that, for any M, χ(a)χ(b)
• ηa

,κ, OpW (f ◦ Φt ǫ)ηb ,κ

• can be

written

• |b−a−n|=O(r)

(−1)Nnq·npeiω(κ,n)+iω(n,b)/2 ηa

, OpW (f ◦ Φt ǫ)ηb−n

• L2 + O(M)

uniformly with respect to a, b ∈ [0, 1)2d. Now, if d = 1 and 5σ < τ, using (4.5) and proceeding similarly to the proof of Theorem 4.3, we see that for ǫ small enough and Γ < ΓA sufficiently close to ΓA

• ηa

, OpW (f ◦ Φt ǫ)ηb−n

• L2 −
• ηa

, ηb−n

• L2
• T2 f(x)dx = e−tΓO(−1/2 + r/)

uniformly on the set where |b − n − a| = O(r), a, b ∈ [0, 1)2. This shows that (2π)−2

• T2
• T2 λ(a)λ(b)
• ηa

,κ, OpW (f ◦ Φt ǫ)ηb ,κ

• dadb − ||ψ,κ||2
• T2 f(x)dx = O(e−tΓ−1/2−5σ)

and the result follows.

• A

A mixing theorem for perturbations of hyperbolic maps

• n T2.

Let A be a 2 × 2 matrix with integer entries such that |trA| > 2 and detA = 1. For notational convenience, we assume that its eigenvalues are positive and we note them e±ΓA, with ΓA > 0. Let φǫ be a measure preserving diffeomorphsim on T2, depending on a parameter ǫ, such that φǫ → id in C3(T2) as ǫ → 0. We define the associated Ruelle-Perron-Frobenius operator Lǫ as the map Lǫg := g ◦ T −1

ǫ

, Tǫ := φǫ ◦ A. Using [3] (more precisely (2.1.7), Example 2.2.6 and Theorem 3) one has the following result. 20

Theorem A.1 ([3]). For any Γ < ΓA, one can find ǫ0 > 0 small enough such that the following property holds: for all |ǫ| ≤ ǫ0, there exists a Banach space Bǫ of distributions of order 1, containing C1(T2), with norm ||.||ǫ such that ||g||ǫ ≤ Cǫ||g||W 1,1, ∀ f ∈ C1(T2) (with ||g||W 1,1 =

• T2 |g| +
• T2 |∇g|) and such that

Lǫ = Π1 + Rǫ with RǫΠ1 = Π1Rǫ = 0 where Π1g = g, 11 and Rǫ is a bounded operator on Bǫ with spectral radius lower than e−Γ. Here ., . is the pairing between distributions of order 1 and C1 functions. As a direct consequence, we obtain Corollary A.2. For all Γ < ΓA, there exists ǫ0 such that, for all |ǫ| < ǫ0, one can find Cǫ,Γ satisfying

• T2 f
• T t

ǫ (x)

• g(x)dx −
• T2 f
• T2 g
• ≤ Cǫ,Γe−tΓ||f||C1||g||W 1,1,

for all f, g ∈ C1(T2), t ≥ 0.

B Generalized coherent states decompositions

In this appendix, we briefly recall some results on coherent states decompositions as well as some convenient tools for the proof of Lemma 4.8. As it is for instance proven in [14], it is well known that for any u ∈ S(Rd) one has u = (2π)−d

• R2d ϕa

, uL2 ϕa da

(B.1) where ϕa

is defined by (4.2) with µ = 1/2. This implies in particular that, for any ˜

ϕ ∈ S(Rd), ||u||2

L2|| ˜

ϕ||2

L2 = (2π)−d

• R2d | ˜

ϕa

, uL2|2 da.

(B.2) This decomposition on L2(Rd), known as the coherent states decomposition especially when ϕ(q) = η(q) = π−d/4e−q2/2, gives rise to a decomposition on H(κ) S(κ)u = (2π)−d

• T2d
• ϕa

,κ, S(κ)u

• H(κ) ϕa

,κ da,

(B.3) with the notation of (4.3). This is proven in [7]. Note the important consequence of that formula: for any ˜ ϕ ∈ S(Rd) (2π)−d

• T2d |S(κ) ˜

ϕa

, S(κ)u|2 da = C ˜ ϕ ||S(κ)u||2 H(κ) ,

∀ u ∈ S(Rd). (B.4) These decompositions are particularly convenient since one knows rather precisely the action of pseudodifferential operators on functions of the form (4.2), as we shall see in Lemma B.1 below. Motivated by Lemma 4.8, we shall consider functions f depending possibly on . Let ε > 0 and assume that r is a sequence such that r ≥ 1/2−ε and let f be a family of functions in B(R2d) such that

• ∂γf (x)
• ≤ Cγr−|γ|
• ,

x ∈ R2d. (B.5) 21

Lemma B.1. There exists a family Pγ of differential operators with polynomial coefficients (inde- pendent of ) such that for any f as above and any M > 0, there exists symbols f (,M,γ) satisfying (B.5) as well and differential operators QM

γ with polynomial coefficients (independent of too) such

that OpW (f )U(a)ϕ =

• |γ|<M

|γ|/2∂γf (a)U(a)(Pγϕ) + Mε

• |γ|≤2M

OpW (f ,M,γ)U(a)(QM

γ ϕ).

Whenever A = Pγ or QM

γ , we have set (Aϕ)(q) = h−d/4(Aϕ)(q/1/2).

• Proof. It is essentially standard. Since U(−a)OpW (f)U(a) = OpW (f(. + a)), we are left with

the case a = 0. Then, the result simply follows by writing the Taylor expansion of f at 0 and integrating by parts.

• Remark. The operators Pγ can be computed explicitly and in particular P0 = I.

Combining this result and (4.4), it is not hard to deduce that for any f ∈ C∞(T2d) satisfying (B.5), one has, for all M > 0,

• OpW (f )ϕa

,κ −

• |γ|<M

|γ|/2∂γf (a)S(κ)U(a)(Pγϕ)

• H(κ)

≤ CMε (B.6) uniformly with respect to a ∈ [0, 1)2d. We are now ready for the proof of Lemma 4.8. Proof of Lemma 4.8. We only have to show the existence of a sequence r ≥ 1/2−ε for some ε > 0, satisfying r → 0, such that, if 0 ≤ χ ≤ 1 is supported close to 0 and ≡ 1 near 0 then χ(a) :=

• n∈Z2d

χ a + n r

• will satisfy the result. Let us fix ε > 0. Then for any sequence r ≥ 1/2−ε, using the Proposition

2.1, one has OpW (1 − χ)2 =

• j<M

jOpW (χj,) + o(1) in operator norm, provided M = M(ε) is large enough. The symbols χj, are such that ∂γχj, = O(r−|γ|−2j

• ) and χ0, = (1 − χ)2, thus using (B.3), (B.4) and (B.6), one has
• OpW (1 − χ)Ψ,κ
• 2 = (2π)−d
• T2d(1 − χ(a))2
• ηa

,κ, Ψ,κ

• 2 da + o(1)

using also the fact that ||Ψ,κ|| → 1. By Taylor formula, there exists a function ˜ χ ∈ C∞(T2d), independent of , such that ˜ χ(0) = 0 and (1 − χ(a))2 ≤ ˜ χ2(a)/r2

. Since

(2π)−d

• T2d ˜

χ(a)2

• ηa

,κ, Ψ,κ

• 2 da → 0

(B.7) by (4.20) applied to f = ˜ χ2, we see that ||OpW (1 − χ)Ψ,κ|| → 0 provided r2

→ 0 more slowly

than the left hand side of (B.7). Furthermore there is no restriction to choose r ≥ 1/2−ε. Finally, we remark that OpW (χ)Ψ,κ = (2π)−d

• T2d
• ηa

,κ, Ψ,κ

• χ(a)ηa

,κ da + o(1)

22

by (B.3), (B.4) and (B.6) again which completes the proof of (4.21). For the converse, we note that

• Ψ,κ, OpW (f)Ψ,κ
• − (2π)−d
• T2d χ(a)
• ηa

,κ, Ψ,κ

Ψ,κ, OpW (f)ηa

,κ

• da → 0.

The result follows then easily from the dominated convergence theorem using (B.6) and (B.4).

References

[1] N. Anantharaman, The eigenfunctions of the Laplacian do not concentrate on sets of small topological entropy, preprint june 2004. [2] M. Brin, G. Stuck, Introduction to dynamical systems, Cambridge Univ. Press (2002). [3] M. Blank, G. Keller, C. Liverani, Ruelle-Perron-Frobenius spectrum for Anosov maps, Nonlinearity 15, no. 6, 1905-1973 (2002). [4] F. Bonecchi, S. De Bi` evre, Exponential mixing and | ln | time scales in quantized hyper- bolic maps on the torus, Comm. Math. Phys. 211, 659-686 (2000). [5] , Controlling strong scarring for quantized ergodic toral automorphisms, Duke Math. J, Vol. 117, No. 3, 571-587 (2003). [6] , Controlling strong scarring for quantized ergodic toral automorphisms, Section 5, mp arc 02-81 (2002). [7] A. Bouzouina, S. De Bi` evre, Equipartition of the eigenfunctions of quantized ergodic maps

• n the torus, Comm. Math. Phys. 178, 83-105 (1996).

[8] A. Bouzouina, D. Robert, Uniform semiclassical estimates for the propagation of quantum

• bservables, Duke Math. J, Vol. 111, No. 2, 223-252 (2002).

[9] Y. Colin de Verdi` ere , Ergodicit´ e et fonctions propres du Laplacien, Commun. Math. Phys. 102, 497-502 (1985). [10] Y.V. Egorov, On canonical transformations of pseudo-differential operators (in Russian), Uspekhi Mat. Nauk. 24, no. 5, 235-236 (1969). [11] F. Faure, S. Nonnenmacher, S. De Bi` evre, Scarred eigenstates for quantum cats of minimal periods, Commun. Math. Phys. 239, 449-492 (2003). [12] F. Faure, S. Nonnenmacher, On the maximal scarring for quantum cat map eigenstates,

• Commun. Math. Phys. 245, 201-214 (2004).

[13] F. Faure, S. Nonnenmacher, contribution at the Workshop on Random Matrix theory and Arithmetic Aspects of Quantum Chaos, Newton Institute, Cambridge, june 2004. [14] G.B. Folland, Harmonic analysis in phase space, Ann. Math. Studies, Princeton Univ. Press 122, (1989). [15] B. Helffer, A. Martinez, D. Robert, Ergodicit´ e et limite semi-classique, Comm. Math. Phys 109, 313-326 (1987). 23

[16] L. H¨

• rmander, The analysis of linear partial differential operators III, Springer-Verlag

(1985). [17] E. Lindenstrauss, Invariant measures and arithmetic quantum unique ergodicity, Annals of Math., to appear. [18] D. Robert, Autour de l’approximation semi-classique, Progress in mathematics, 68, Birkha¨ user (1987). [19] Z. Rudnick, P. Sarnak, The behaviour of eigenstates of hyperbolic arithmetic manifolds,

• Commun. Math. Phys. 161, 1, 195-213 (1994).

[20] A. Schnirelman, Ergodic properties of eigenfunctions, Usp. Math. Nauk. 29, 181-182 (1974). [21] R. Schubert, Semiclassical behaviour of expectation values in time evolved coherent states for large times, preprint january 2004. [22] M. Taylor, Pseudo-differential operators, Princeton Mathematical Series 34, Princeton Uni- versity Press (1981). [23] F. Tr` eves, Introduction to pseudo-differential and Fourier integral operators, Vol. 2: Fourier integral operators, Univ. Ser. Math., Plenum, New-York (1980). [24] Zelditch, Uniform distribution of the eigenfunctions on compact hyperbolic surfaces, Duke

• Math. J 55, 919-941 (1987).

24